A lobster ate the interface
The most interesting product of the last year wasn't shipped by a product team. OpenClaw, the scrappy open-source agent that started life as a weekend project with a lobster for a mascot, went viral for one simple reason: it doesn't chat with you. It does things for you.
It lives on your machine. It talks to you through the apps you already use. You tell it "book the table, reply to mum, keep an eye on that price." Instead of giving you a nicely formatted paragraph about how one might do that, it goes and does it. People didn't fall in love with its design. There barely is one. They fell in love with the absence of work.
From answers to actions
For three years we've been calling chatbots "AI products." But a chat window is a vending machine for text: you put a prompt in, you get words out, and then you still have to do the thing. The chat window was training wheels: a familiar metaphor to ease us into talking to machines.
Agents take the training wheels off. The unit of value is no longer the answer; it's the completed task. That flips a fifteen-year-old assumption in our discipline: that our job is to make software easier to operate. Nobody wants to operate software. They want outcomes. The interface was always overhead. We just had no alternative to offer.
I've spent my career designing how people talk to technology. The new brief is designing how technology acts on our behalf, and stays worthy of that trust.
The new design problems
If you think agents make designers obsolete, look closer: every hard problem OpenClaw surfaced is a design problem wearing a security costume.
Permission design. What may the agent do alone, what needs a nod, what should it never touch? "Grant full access" is the new "accept all cookies": a consent flow nobody designed.
Legibility of action. When a human assistant does something for you, you can ask what they did and why. An agent needs the same: a record you can read, and an undo you can trust. Reversibility is the new usability.
Knowing when to interrupt. An agent that asks about everything is a chatbot with extra steps. One that never asks is a liability with your credit card. Calibrating that line (per person, per task, per stake) is interaction design at its purest.
Failure recovery. Agents fail weirdly. They half-do things. Recovering gracefully from a half-booked trip is a flow someone has to design, because the model certainly won't.
The uncomfortable part
OpenClaw's virality came with stories of agents running with far too much access, doing far too much, far too confidently. Handing a process the keys to your digital life is a UX problem disguised as a security one: people will always trade safety for convenience if the convenience is real and the risk is abstract. Our job is to make the risk legible before the trade, not after.
In summary
Every platform shift looks like a toy first. A voice-based social network in 2013 looked like a toy. A lobster-branded agent in someone's terminal looks like a toy. Then one day you notice the toy is answering your messages, and the question is no longer whether agents will mediate our digital lives. It's whether the people designing that mediation will think about the humans on both ends of it.
The interface of the next decade isn't a screen. It's a colleague. Design accordingly.